Comparison of Uncertainty Quantification with Deep Learning in Time Series Regression

Increasingly high-stakes decisions are made using neural networks in order to makepredictions. Specifically, meteorologists and hedge funds apply these techniquesto time series data. When it comes to prediction, there are certain limitations formachine learning models (such as lack of expressiveness, vulnerability of domainshifts and overconfidence) which can be solved using uncertainty estimation. Thereis a set of expectations regarding how uncertainty should “behave". For instance, awider prediction horizon should lead to more uncertainty or the model’s confidenceshould be proportional to its accuracy. In this paper, different uncertainty estimationmethods are compared to forecast meteorological time series data and evaluate theseexpectations. The results show how each uncertainty estimation method performson the forecasting task, which partially evaluates the robustness of predicteduncertainty